test(where-matcher): empty the combinator-blind ledger — 75 files, refuse unrecognised $-keys

[os-zhuang]

Closes#8582

What

Sweeps all 75 files (78 matchers + 1 unjudged) grandfathered in scripts/where-matcher-conformance.baseline.json — test-double matches(row, where) fakes with no combinator branch at all, which read $or/$and as an ordinary field name, compared row.$or (undefined) against the array, and silently dropped every row a combinator query touched.

Default remedy (per the issue): make the double refuse — throw on an unrecognised $-prefixed key, the practice already pinned in packages/objectql/src/engine-autonumber-*.test.ts. Implemented the combinator instead only where a suite genuinely exercises it (see below).

pnpm check:where-matcher (final state):

✓ where-matcher conformance holds: 168 matcher(s) discovered, 168 answer the combinator battery
correctly or refuse it loudly (82 refuse).
0 silently-wrong and 0 unjudged matcher(s) in 0 grandfathered file(s); none new.

The ledger is now {} — empty, matching failure-shape-(a)'s steady state.

Full file surface (cross-domain exception, per the claim comment)

75 test files across metadata-protocol, objectql, plugin-approvals, plugin-auth, plugin-email, plugin-hono-server, plugin-reports, plugin-security, plugin-sharing, rest, service-analytics, service-automation, service-messaging, service-queue, service-settings, service-storage, trigger-schedule — test doubles only, zero production files. Disjoint from the two other in-flight production-file claims this batch (#8559packages/objectql/src/engine.ts, #8460packages/objectql/src/registry.ts) — neither touched here.

Commits, per-package batches with the ledger lowered in the same commit each time (self-proving against the gate):

1. metadata-protocol, objectql, plugin-approvals, plugin-auth (39 files)
2. plugin-email, plugin-hono-server, plugin-reports, plugin-security, plugin-sharing, rest (22 files)
3. service-analytics, service-automation, service-messaging, service-queue, service-settings, service-storage, trigger-schedule (14 files)
4. One follow-up fix (below)

Two files needed more than the mechanical guard

• packages/plugins/plugin-sharing/src/business-unit-graph.test.ts already implements $or correctly (a real, exercised recursive conjunction). Added a refusal for any other$-key ($and included) rather than growing a second combinator implementation nothing in the suite exercises.
• packages/services/service-analytics/src/__tests__/objectql-daterange.test.ts — the blanket refusal I first landed (batch 3) turned a passing test red: query() drives through fieldLeaves's real NULL-safe $ne expansion (非否定路径上的 $ne / $nin / $notContains:driver-sql 排除 NULL 行,driver-memory / formula 返回它们(#5146 只裁定了 $not) #5298), which emits $and: [{ $or: [{field:null},{field:{$ne}}] }] — this double genuinely receives $or at runtime, it is not dormant here. Per the issue's own guidance ("option 1 only where a suite genuinely exercises combinators"), implemented $or the same way the file already implements $and (a .some() over the branches) instead of refusing it. Caught by running the full test suite before reporting, not by the gate (the gate only asks the four synthetic battery questions against the lifted function in isolation — it has no way to see what a real caller in this file actually hands the matcher).

The 1 unjudged matcher

packages/plugins/plugin-email/src/outbox-sweep.test.ts. Its nested-operator throw message (`fakeEngine: unsupported operator ${op}`) coincidentally repeats its enclosing factory function's exact name (function fakeEngine(seed: RowSeed[])), and the gate's identifiersIn() text-scanner reads that string-literal word as if it were a reference to the factory — pulling the whole closure (and its unsuppliable seed parameter) into the standalone extraction, which then threw ReferenceError: seed is not defined no matter how the retry/drop loop iterated.

Disposition: RESTRUCTURE, not extractor-reach — renamed the message so it no longer shadows the enclosing scope (one line, zero blast radius) rather than changing the shared identifiersIn() text-scanner every one of the 168 discovered matchers depends on. The matcher now judges CONFORMING (refuses).

Out-of-scope finding

Filed #8615: the gate's discovery heuristic requires a matcher to be a function with two or more identifier parameters, so a single-param .filter((r) => Object.entries(where).every(...)) closure that captures where/filter from an enclosing scope is structurally invisible to it — 17 such sites across 16 files carry the identical shape-(b) pattern, completely ungraded by check:where-matcher. Not fixed here (outside #8582's worklist, which is scoped to the ledger); left for triage.

Verification

• pnpm check:where-matcher — green (0 silently wrong, 0 unjudged), see above.
• pnpm test scoped to all 17 touched packages (metadata-protocol, objectql, plugin-approvals, plugin-auth, plugin-email, plugin-hono-server, plugin-reports, plugin-security, plugin-sharing, rest, service-analytics, service-automation, service-messaging, service-queue, service-settings, service-storage, trigger-schedule) — all green, real numbers:
  • metadata-protocol 85/85 files, 1261/1261 tests
  • objectql 202/202 files, 3559/3559 tests
  • plugin-approvals 21/21 files, 458/458 tests
  • plugin-auth 53/53 files, 1195/1195 tests
  • plugin-reports 3/3, 70/70 · plugin-hono-server 18/18, 211/211 · plugin-sharing 21/21, 569/569
  • service-queue 3/3, 53/53 · service-messaging 20/20, 219/219 · service-settings 23/23, 451/451
  • service-analytics 75/75, 1675/1675 (after the $or fix) · service-storage 24/24, 361/361
  • plugin-email 23/23, 369/369 · plugin-security 55/55, 1079/1079
  • rest 114/114, 1881/1881 · service-automation 80/80, 955/955 · trigger-schedule 4/4, 46/46
• pnpm typecheck on every touched package that declares its own typecheck script — clean (service-analytics, service-automation, service-storage carry no typecheck script; pre-existing DEBT/EXEMPT-ledger packages, untouched by this test-only diff).
• Derived gates (node scripts/pm/dispatch-gates.mjs <changed paths>), all green: check:cross-package-test-inputs, check:durability-log-level, check:test-source-alias, check:type-source-resolution, check:engine-split-ratio, check:query-options-erasure, check:type-check-coverage (436 frozen DEBT / 1533 frozen TEST_DEBT — unchanged from main, no regression), check:i18n (9 packages, all in sync), check:nul-bytes.

Not in scope

• No shared matchesWhere extraction — ruled NO on finding: 16 test doubles short-circuit $or in their WHERE matcher, dropping sibling filters — the suite stays green while testing a different query #7620, not re-litigated.
• Test-only surface — no changeset (skip-changeset label applied).
• #8559 / #8460's production files (packages/objectql/src/engine.ts, packages/objectql/src/registry.ts) — not touched.

---

Generated by Claude Code

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

1 Skipped Deployment

Project	Deployment	Actions	Updated (UTC)
objectstack	Ignored		Aug 14, 2026 2:33am

[github-actions]

📓 Docs Drift Check

No hand-written docs reference the 0 changed package(s). ✅

[os-zhuang]

PM review (domain:engine-core seat, #6019): ACCEPT. Ready + auto-merge (squash). All 35 checks green or skipped, including check:where-matcher itself and all three Test Core shards.

⭐ Two things here are better than the card asked for, and both came from running the real suites rather than trusting the gate's synthetic battery.

1. service-analytics/objectql-daterange.test.ts genuinely receives an $or — production fieldLeaves' NULL-safe $ne expansion hands it one. The gate's four-probe battery could not have told you that; only the real suite did. So it correctly got a real $or implementation instead of the default throw. ⚠️ That is the standing caveat on this whole remedy: "refuse" is right only where the combinator never arrives, and the suite, not the gate, is what decides that.
2. The 1 unjudged matcher was a detector bug, not a double bug. Its throw-message text coincidentally repeated its enclosing factory's exact name, so the gate's text-scanning identifiersIn() pulled in an unsuppliable closure parameter. Fixed by restructuring the message — ⛔ correctly not by touching the shared extractor.

⚠️The empty ledger ({}) must not be read as "the repo is swept clean." The same sweep filed #8615: the discovery heuristic misses single-param inline filter closures (.filter((r) => Object.entries(where).every(...)) capturing where from an enclosing scope), and 17 sites across 16 files carry the identical shape-(b) pattern, invisible to this gate. So 0 silently-wrong in 0 grandfathered files is a true statement about what the gate can see, and the gate's field of view is now the open question. That is exactly the "the instances are fixed, nothing holds them fixed" shape one level up again — ⛔ triage's to grade, not this seat's, but it belongs in the record next to the empty ledger.

Cross-domain conditions met: full file surface declared, 17 packages re-run green, ledger lowered in the same commit so the PR self-proves.

---

Generated by Claude Code